Doublespeed

Create multiple posts in bounded parallel batches, with up to 50 objects using the same fields as create_post. Each item defaults to scheduled and can publish to its connected account; use status: draft to save without scheduling. Concurrency defaults to 3 and can be set with batch_size. A completed response reports each item's success or failure. This is not an atomic operation: some posts may be created before another fails or the request times out. Check for already-created posts before retrying an interrupted batch.

create_posts_bulk

Create a shareable review link for one draft, or return its existing link. Pass the draft `group_id`. If existing is true, the supplied password and allow_comments settings are not applied; this call does not confirm the existing link's protection. A new link without a password is accessible to anyone who has its URL.

create_review_link

Create a v2 scene template for the active product. For a simple named draft, pass metadata.name (and optional metadata.prompt) and omit scene_data; the app creates a blank slide. Only send scene_data when you already have a full preview from render_slides. Omit thumbnail_url and reference_import unless you have real values. Call once; do not retry on failure. Returns the new template id; use get_template to inspect it or list_templates to rediscover it later.

create_template

Render slideshow sceneData to slide image URLs for visual inspection. Returns slide images and hosted URLs when hosting succeeds. Does not save a draft. Supports visual iteration on slide layouts, including 1080x1920 vertical slides. Use absolute pixel coordinates for scene block layout. Style every text block by naming a preset: `style.stylePreset` set to a name from `list_style_presets` (TikTok-style slides commonly use `tiktok-stroke` or `tiktok-bg`). The renderer expands the name into the full style using the same expansion as the draft write path. Do not hand-set fontFamily, fontSize, lineHeight, letterSpacing, padding, borderRadius, or backgroundShape alongside a preset; color and bounds are the fields meant to be tuned per slide. If the preview comes back with cramped letters, overlapping lines, or unreadable text, fix the sceneData and render again before presenting it as final. To render several independent decks (for example one per variant) in a single call, pass `scenes` as an array of 1 to 8 `{ name?, scene_data }` entries, or pass top-level `scene_data` (exactly one of the two). Results come back per scene, labelled by `name`, and one scene's failure does not fail the others. The same payload still includes top-level `slides` / `hostedUrls` for callers that read those fields. Treat `warnings` as unreadable slides and re-render before shipping. Keep every text block inside the platform safe area: on a 1080x1920 slide that is x 46-1033, y 115-1545, minus the platform UI sitting inside it (x 313-762, y 133-216; x 887-1019, y 960-1092; x 886-1020, y 1152-1316; x 887-1019, y 1358-1521). Text crossing those rects is covered by the app's own chrome once the slide is posted, so it reads as a broken slide however good the copy is. A block grows to fit its text, and a centered block grows around its center, so leave room on both sides: wide copy belongs above the action rail rather than beside it. The write boundary moves a block down or up when that is enough to clear the chrome, but it never rewrites or shrinks copy. Size the text to the room it has. Example `sceneData` payload for `render_slides` (1080x1920, all layout values in pixels): ```json { "id": "scene-1772623220885", "name": "TikTok slide preview", "pages": [ { "id": "page-1772623220885", "name": "Page 1", "width": 1080, "height": 1920, "background": "#ffffff", "blocks": [ { "id": "background", "type": "background", "bounds": { "x": 0, "y": 0, "width": 1080, "height": 1920 }, "rotation": 0, "zIndex": 0, "style": { "type": "color", "color": "#ffffff" } }, { "id": "image-1772623222845", "type": "image", "bounds": { "x": 0, "y": 0, "width": 1080, "height": 1920 }, "rotation": 0, "zIndex": 1, "src": "https://cdn.example.com/background.jpg", "style": { "objectFit": "cover", "borderRadius": 0, "opacity": 1 } }, { "id": "text-1772623231625", "type": "text", "bounds": { "x": 140, "y": 300, "width": 800, "height": 220 }, "rotation": 0, "zIndex": 2, "text": "example text", "style": { "stylePreset": "tiktok-stroke", "verticalAlign": "middle" }, "autoResize": true, "sceneWidth": 1080 } ] } ], "pageOrder": ["page-1772623220885"], "activePageId": "page-1772623220885", "defaultPageSize": { "width": 1080, "height": 1920 }, "templateMode": true } ```

render_slides

Edit a variant on a draft using either a review link `token` or by draft `group_id` (exactly one). Provide either one `variant_id` or several `variant_ids`. Targeted block helpers preserve fields you do not specify; scene_data replaces the full scene. The response reports applied edits and errors; inspect it before retrying or making further changes. Edit paths: 1. Leaf fields (caption, accountId, musicLink/musicUrl, muteMusic, status) - safe to patch independently, no baseRevision needed. 2. textBlockEdits - edit existing on-screen overlay text on video or slideshow variants. Send targeted edits by blockId from the draft's textOverlays. No baseRevision needed. 3. textBlockInserts - ADD on-screen overlay text to a hook-demo video segment. Each entry needs `segment_index` and `text`. The text fills the segment's primary (first empty) text column using the site's default TikTok-style format, so it shows in BOTH the bulk-video table editor and the review link; a new column is only appended when every existing text column is already filled. Use this when you need to ADD text that does not yet exist (textBlockEdits cannot create blocks). No baseRevision needed. 4. blockEdits - change ANYTHING about a slideshow block: text, image, font size, colour, alignment, line height, letter spacing, stroke, background, padding, position, size, rotation, stacking. Send only what changes; style and bounds are shallow merged so nothing else is touched. No baseRevision needed. 5. emphasisEdits - style individual WORDS inside a slideshow text block (bold, colour, size, font) while the rest of the paragraph keeps the block style. Pick the words with a rule, never offsets: {"type":"uppercase_words"} for every fully capitalised word, or {"type":"words","words":[...]} for named ones. The rule resolves against each variant's own copy, so one call with variant_ids styles every variant at once. `clear: true` removes per-word styling. No baseRevision needed. 6. segmentEdits - video on-screen text and its styling, image overlays, video filters, framing, crop, canvas colour, per-clip trim and speed. Video variants only. No baseRevision needed. 7. structureEdits - add or remove blocks, and duplicate, delete or reorder slideshow pages. Positional variableKeys and pageOrder are repaired for you. Slideshow variants only. No baseRevision needed. 8. imageBlockEdits - swaps a slideshow image src only. 9. sceneData - full structure replace when a complete scene rewrite is required. Requires baseRevision when using a token; ignored for group_id. scene_data is mutually exclusive with all targeted edit helpers: text_block_edits, text_block_inserts, image_block_edits, block_edits, emphasis_edits, segment_edits, and structure_edits. Targeted helpers can be combined. `status` is review-link-only (drafts without a share link have no reviewer status).

patch_draft

Transcribe a publicly fetchable HTTP(S) audio or video URL with Whisper and return word-level timings ({ word, start, end } in seconds). Feed the words to save_editor_project as captions to get editable, word-timed caption items on the editor timeline. For generated voiceovers, use the hosted audioUrl from generate_voiceover; data URI fallbacks are not supported.

generate_captions

Submit an async video generation job. Returns a media file ID to poll with check_generation_status. Videos are generated via the configured model provider and cost credits. Call list_models to see supported models, parameters, and reference pricing; the actual charge can depend on product-specific pricing rules.

generate_video

Generate a voiceover audio track from a script using ElevenLabs text-to-speech. Returns audioUrl and an estimated duration in milliseconds. The URL is normally hosted; if hosting fails, it can be a data URI, which generate_captions cannot accept. Use a hosted HTTP(S) URL for caption generation and editor workflows that fetch remote audio. Omit voice_id to use the default narrator voice; call with list_voices=true to see the available voices first.

generate_voiceover

Get details of a specific social media account including username, bio, status, platform, and profile info.

get_account

Get details of the currently active product including title, overview, brand guide, target audience, and other context.

get_product

Fetch a draft's current state. Two modes: 1. Summary (default): compact overview of all variants with textOverlays for quick scanning. No segments or sceneData. 2. Detail (variant_index or variant_id provided): full segments, sceneData, and textOverlays for one variant. Summary mode includes addressable blocks. Slideshows: blocks (every addressable block with its current text/src, style and bounds, a variableKey that stays stable across variants, and spans for any per-word styling already applied). Videos: videoSegments (text blocks with their flat styling, image overlays, filters, framing and clips). Plus the narrower textOverlays and imageBlocks views. A block's style applies to the WHOLE paragraph. Per-word colour, bold, size, or font lives in spans on that block, not as a second overlapping block. Detail mode returns the full segments or sceneData for one variant.

get_draft

Get exact aggregate post performance for the active product across the full filtered post set in one call. Use this for totals, averages, medians, maximums, engagement, or per-account performance questions. No pagination or sampling is needed.

get_post_metrics

Returns your identity, permissions, and the currently active product for this MCP session.

whoami

Copy a publicly downloadable media file (direct mp4/mov/webm/jpg/png/webp URL) into a product media folder and return its durable media URL. Use this when a product needs a stored copy of externally hosted media. The URL must serve the file bytes directly: page URLs such as YouTube or TikTok watch pages are not downloaded here. With probe_only: true it reads remote media bytes to inspect metadata (kind, width, height, duration) without saving an imported copy; use that before save_editor_project so scene trims stay inside the source and contain/blur-fill fit modes know the source aspect (pass the result as source_duration_ms, source_width, source_height).

import_media_url

List posts for a specific account within the active product. Supports pagination, status filtering, and sort direction. Returns summary rows by default; heavy fields (scene_data, template_data, image_urls) require explicit include_* flags. Same row semantics as list_posts: `succeeded` = live, `post_time` is the requested slot (use `published_at` for when it went live), and `public_url_status: pending_link` means live but the public URL is not linked yet. Confirmed publication data stays available during `pending_private`/`pending_delete` until removal completes.

list_account_posts

List social media accounts for the active product. Supports pagination, status filtering, and search by username.

list_accounts

List supported image and video generation models, input requirements, capabilities, and default pricing. Prices are reference defaults; product-specific pricing rules and billing rounding may change the actual charge.

list_models

Show the order the active product's queued posts will go out in, per account, with each post's expected publish time. Every scheduled post takes the account's next open slot from the scheduler (posting windows, cooldown, daily target, device gaps); this is where to see which slot a post got. `scheduled_at` is the locked publish time (ISO 8601, UTC) or null while the post is queued without one (`queue_state: queued`); `estimated_at` then carries the time it would get if it took a slot now, from the account's own windows and daily target. `position` is 1-based firing order within the account. `reorderable: true` marks posts whose slots reorder_posting_queue may swap. Posts mid-publish (pending) have left the queue; use list_posts for them. `truncated: true` means the product holds more queued posts than one listing returns; narrow with account_id.

list_posting_queue

List posts for the active product. Supports pagination, status filtering, date range, sorting, and search. Returns summary rows by default; heavy fields (scene_data, template_data, image_urls) require explicit include_* flags. Reading a row: `status` is the posting state (`succeeded` = live on the platform; `scheduled`/`pending`/`need_approval` = still queued; `failed`; `draft`). `post_time` is the requested queue slot, not when the post went live; use `published_at` (ISO, null until live). `public_post_url` is stamped by analytics reconciliation after the post is live and can trail it by hours, so `public_url_status: pending_link` on a `succeeded` row means live-but-not-linked-yet, not stuck. Confirmed publication data stays available during `pending_private`/`pending_delete` until removal completes.

list_posts

List products you have access to, and show which product is currently active for this MCP session. Supports `search` (case-insensitive title match) and `page`/`page_size` pagination: required on admin sessions because the full admin listing can exceed response limits. When several products share a title, each row includes `duplicateCount`, `isPaid` (has a subscription), and `isEmpty` (no subscription, accounts, or posts). Always pass the `product_id` the user named or confirmed.

list_products

List the canonical text style presets used by the editor (font family, weight, size, color, background, stroke, spacing, alignment). Returns each preset's name, a short description, the surfaces it may be used on (video overlay text is limited to tiktok-sans and tiktok-stroke; slideshows may use all four), and its exact style values, plus usage guidance for applying a preset via patch_draft text_block_inserts or upsert_video_draft textBlocks. These are text style presets, not hook_demo_presets. Also returns `fontFamilies`: the built-in families plus this workspace's uploaded custom fonts; any of them is valid in `fontFamily` (block, span, or emphasis).

list_style_presets

List the active product's wiki pages: its reference knowledge (overview, brand and voice, formats, asset locations, decisions, learnings, one accounts/<handle>/persona page per account). Rules are not pages; they live on the product Rules page. Call this once after set_product and read the relevant pages with wiki_get_page before content work. Each entry carries slug, title, visibility, current revision number, generation/review flags, and sections (level-2 headings). Product members edit wiki pages in the dashboard. Staff sessions see the product's own pages plus company-wide (all_products) pages; a client session sees only the pages published to it (visibility client).

wiki_list_pages

Fetch a single v2 template by id, including template.sceneData and its variables. The result card shows its cover thumbnail. To preview a slideshow template's pages, pass template.sceneData as scene_data to render_slides. To save a derived slideshow, pass the edited scene_data and the template id as source_template_id to upsert_slideshow_draft. Video drafts require the videos and hook_demo_state inputs described by upsert_video_draft; it has no source_template_id parameter. Access is scoped to templates owned by the active product, created by the caller, or marked public.

get_template

Prepare a media-library upload for one or more local files. This may create the requested private media folder and ensures the public-read Azure media container exists, then returns a time-limited SAS upload URL and blob path per file. Upload each file's bytes directly to its uploadUrl with an HTTP PUT, then call commit_media_upload (which also accepts a files batch) to register them and get durable media URLs. Pass file_name for a single file or files for a batch.

prepare_media_upload

Create a post from a draft variant. Defaults to scheduled, which queues it for publication to its connected account; set status to draft to save without scheduling. Resolves the variant from a draft (`group_id`) or review link (`token`) and queues it using ONLY the variant's own data: caption, account, music (including the mute-music toggle set via patch_draft `mute_music`), and media all come from the variant itself. There are no overrides: if a required field is missing on the variant, the call fails with a clear message and the caller must edit the variant (e.g. via patch_draft) before retrying. Video variants are composed the way the bulk-video table's Queue composes them: when the stored video is only the raw source clip (the draft has overlay text or image blocks, clip trims or speed changes, a transform, crop, filter, or more than one segment), the overlays and the music track are rendered into the video server-side and that render is queued; a single-clip variant with no overlays or edits queues its stored `videoUrl` as is. Composing adds render time, which varies with the content and load, and the response then carries `composedVideo`. Slideshow variants are server-rendered from `variant.sceneData` at queue time. The created post is tagged with provenance (sourceVariantId / sourceVersionId / sourceGroupId) so the drafts UI links it back. Cross-product safety: the active MCP product must own the draft, except when the caller is an admin (admins can queue drafts from any product). A slideshow queued with no audio succeeds but returns a `warnings` entry; attach music only when requested or appropriate for the content. Scheduled posts containing #ads are rejected because this posting path cannot complete the advertising disclosure prompt. Keep required disclosures intact and use a posting flow that supports them.

queue_post

Read one wiki page by slug for the active product (its reference knowledge). Read the pages that cover the task (overview first when present, then brand and voice, formats, learnings, the account's accounts/<handle>/persona page) before creating content; rules are not wiki pages, they live on the product Rules page. Returns the page metadata, current revision (send its rev as base_rev when editing), and rendered segments (markdown runs plus managed directive blocks resolved live). A slug not visible in the caller's scope is reported as not found.

wiki_get_page

Render a full video composition through doublespeed's Remotion pipeline. Accepts a Remotion composition payload and returns an mp4 or webm file, or a still frame when `still_at_ms` is set. Pass `project_id` for a saved editor project, or `still_at_ms` to get a single PNG/JPEG frame at that time (single-frame layout check before the full render). Composition timing is millisecond-based: `duration` and every track item's `display.from` / `display.to` must be in milliseconds. For layout, set `details.width`, `details.height`, `details.top`, and `details.left` to match the target canvas. The `text_overlay_data` input is preserved as textOverlayData in the result for downstream review bookkeeping, but it does not render visible text by itself. If you want text burned into the output, include a real `text` track item (or another visual overlay track) inside `trackItemsMap`. Parameter guide: - `composition.duration`: total output length in milliseconds - `composition.fps`: output frame rate, commonly 30 for short-form social video - `composition.size`: final canvas size in pixels, for example `{ width: 1080, height: 1920 }` - `composition.trackItemIds`: ordered track ids to render; later items visually stack above earlier ones - `composition.trackItemsMap[*].display.from` / `.to`: when that track should appear, in milliseconds - `composition.trackItemsMap[*].details.src`: source media URL for image/video/audio tracks - `composition.trackItemsMap[*].details.width` / `.height`: rendered box size on the canvas - `composition.trackItemsMap[*].details.top` / `.left`: rendered box position on the canvas - `text_overlay_data`: review metadata only, useful for bookkeeping. It does not render visible text by itself; include a real text track item for that. After render_video succeeds, do not assume the output is acceptable just because rendering finished. Review pacing, crops, text legibility, overlay placement, transitions, and artifacts before treating the result as final. If you already have structured overlay metadata, pass it as text_overlay_data so it is preserved in the result for downstream review tooling. Example `composition` payload for `render_video`: Important: `duration` and every `display.from` / `display.to` value are milliseconds, not frame counts. ```json { "duration": 6633, "fps": 30, "size": { "width": 1080, "height": 1920 }, "trackItemIds": ["clip-1", "text-1"], "trackItemsMap": { "clip-1": { "id": "clip-1", "type": "video", "display": { "from": 0, "to": 6633 }, "details": { "src": "https://cdn.example.com/source.mp4", "width": 1080, "height": 1920, "top": 0, "left": 0, "objectFit": "cover" } }, "text-1": { "id": "text-1", "type": "text", "display": { "from": 0, "to": 6633 }, "details": { "text": "when i ask\n\"who all gonna be there\"", "fontFamily": "TikTok Sans", "fontSize": 60, "color": "#FFFFFF", "textAlign": "center", "width": 820, "height": 360, "top": 620, "left": 130 } } } } ```

render_video

Render a baked preview video from editor or hook-demo style state. Burns text and image overlays into the returned video from editor-style inputs. Does not persist a draft. Parameter guide: - `clips`: optional timeline clips array when previewing a multi-clip edit - `video_url`: single source video when you do not have a clips array - `video_transform`: placement box for the main video layer in editor pixels - `text_blocks`: editable text overlays that will be burned into the preview video - `image_blocks`: editable image overlays that will be burned into the preview video - `editor_dimensions`: required canvas size object, for example `{ width: 1080, height: 1920 }` for vertical social content - text block styling: set `stylePreset` to `tiktok-sans` or `tiktok-stroke`, the only two presets video overlay text may use (the filled-block presets are slideshow-only), and let the server expand it. Do not hand-set fontFamily, fontSize, fontWeight, lineHeight, letterSpacing, padding, borderRadius, or backgroundShape alongside a preset; color and placement are the fields meant to be tuned. Example `render_video_preview` payload for a hook-demo-style editable preview: ```json { "video_url": "https://cdn.example.com/source.mp4", "editor_dimensions": { "width": 1080, "height": 1920 }, "video_transform": { "x": 0, "y": 0, "width": 1080, "height": 1920 }, "text_blocks": [ { "id": "text-1", "text": "me after deciding\nnot to ask the one\nquestion that\nwouldve ruined the\nwhole night", "x": 140, "y": 560, "width": 800, "height": 380, "stylePreset": "tiktok-stroke", "zIndex": 10 } ] } ```

render_video_preview

Change the order one account's queued posts go out in. Pass the post ids in the order they should fire (first id first). The named posts keep the set of slots they currently hold and get them back in the new order, so the account's posting windows, cooldown, and daily target still hold and posts you do not name keep their times. Only posts list_posting_queue marks `reorderable: true` (scheduled, holding a future slot) can be named; a post firing within the next 5 minutes cannot be moved. Returns the account's queue after the change. The MCP tools cannot pin a publish time (only the posting REST API's scheduledAt can, within the account's scheduling rules); use this to change the order.

reorder_posting_queue

Request deletion (takedown) of a post that is already live or mid-publish (status succeeded, posted, or pending). The operations team is notified and removes the post from the platform; a processed takedown cannot be undone. The post is not deleted immediately: it flips to pending_private and shows as delete-requested in the app until the takedown is processed. Draft and scheduled posts are not eligible for takedown. This tool never hard-deletes anything. If a deletion request is already open for the post, the call returns an error saying so and no duplicate request is created.

request_post_deletion

Create or update an editable video draft backed by bulk-video table (hook-demo) state. Persists editable text, segments, and overlays in the bulk-video table (/{productId}/bulk-video/table). Optional `create_share_link: true` returns a review URL in the same call. Parameter guide: - `group_id`: omit to create a new draft, include to update an existing draft (variant_group) - `videos`: review rows for the draft. For editable reviews, this is the base video that the review page plays behind the saved overlays, not a baked preview file - `hook_demo_state`: required editable source of truth for rows, segments, text blocks, and image blocks - `hook_demo_state.table_rows[*].segments[*].externalVideoUrl`: the source clip URL that the bulk-video table should reopen on edit - `editor_dimensions`: optional editor canvas size used to normalize editor-pixel text/image blocks into hook-demo's stored 0-1 coordinate system - text blocks accept `stylePreset`, which on a video must be `tiktok-sans` or `tiktok-stroke`; those are the only two presets allowed on video overlay text, and any other name is rejected (the filled-block presets are slideshow-only). The preset fills in style fields and any explicit field on the block overrides it. Name the preset; hand-writing fontFamily, fontSize, fontWeight, lineHeight, letterSpacing, padding, borderRadius, or backgroundShape is how drafts end up matching no preset. Tune color and placement per block, and use the same preset across every segment of one video. - `create_share_link`: set to `true` when you want `reviewUrl` and `shareLink` returned immediately - `password`: optional password for a newly created review link when `create_share_link` is true; does not change an existing link's password This tool will normalize editor-style pixel blocks into hook-demo's persisted normalized block schema before saving. Example `upsert_video_draft` payload that keeps the video editable in the bulk-video table: ```json { "draft_name": "saintdairy internal fight walk", "editor_dimensions": { "width": 1080, "height": 1920 }, "create_share_link": true, "videos": [ { "video_url": "https://cdn.example.com/source.mp4", "caption": "me after deciding not to ask the one question that wouldve ruined the whole night", "segments": [ { "videoUrl": "https://cdn.example.com/source.mp4", "textBlocks": [ { "id": "text-1", "text": "me after deciding\nnot to ask the one\nquestion that\nwouldve ruined the\nwhole night", "x": 140, "y": 560, "width": 800, "height": 380, "stylePreset": "tiktok-stroke" } ] } ] } ], "hook_demo_state": { "segment_count": 1, "table_rows": [ { "id": "row-1", "segments": [ { "id": "segment-1", "externalVideoUrl": "https://cdn.example.com/source.mp4", "textBlocks": [ { "id": "text-1", "text": "me after deciding\nnot to ask the one\nquestion that\nwouldve ruined the\nwhole night", "x": 140, "y": 560, "width": 800, "height": 380, "stylePreset": "tiktok-stroke" } ] } ] } ] } } ```

upsert_video_draft

Assemble scene clips, motion clips (self-contained HTML+GSAP animation bundles), native text items, an optional voiceover/music track, and optional word-timed captions into an editor project saved server-side, and return an editor link. Every layer stays fully editable on the /editor timeline: each scene is its own video item (source trim, timeline placement, and canvas box each set independently), each motion clip is its own layered motion item, text items are the editor's own text blocks, audio tracks are audio items, and captions are caption items. Motion clip HTML contract: one self-contained document (inline CSS/JS, hosted asset URLs allowed; a gsap <script> CDN tag is fine), a paused seekable GSAP timeline assigned to window.__motion, window.__motionReady = true once built, deterministic output for any seeked time (use fromTo tweens, no unseeded randomness or wall-clock time), and a transparent document background unless the clip is itself the background layer, so clips overlay like transparent PNGs. Typical flow: upload or import the source (import_media_url probe_only for its duration and size) -> generate_captions on the source -> save_editor_project with scenes trimmed out of the source and captions timing=source -> render_video with project_id. Pass project_id to overwrite an existing project; use patch_editor_project to change single items without resending everything. The response is lean (ids and an item index); pass include_composition=true to also get the full composition back.

save_editor_project

Create or update a slideshow draft backed by sceneData: save the slideshow so the user can review, edit, and post it. Persists slideshow sceneData into the draft system (variant_group). Optional `create_share_link: true` returns a review URL in the same call. Does not render slide images. Parameter guide: - `scene_data`: the normal single-draft path; persist the slideshow structure to save - `variants`: advanced multi-variant slideshow input when you want one draft containing several slideshow variants - `group_id`: omit to create a new draft, include to update an existing slideshow draft (variant_group) - `source_template_id`: id of the template the slideshow started from (`list_templates` / `get_template`); pass it whenever the sceneData began as a template so lineage is preserved. Agent runs are held to this: when the product has templates, a new draft without `source_template_id` is rejected, and the persisted sceneData must be the referenced template's sceneData edited in place, not a rebuilt lookalike. Enforced per text block: keep the template's fontFamily, fontWeight, fontSize, lineHeight, and letterSpacing exactly as fetched; change only text content and image sources, and tune only color and bounds per slide. - text block styling: set `style.stylePreset` to a canonical name from `list_style_presets` and let the server expand it. Any field you also set overrides the preset, so restating the preset's own values (fontFamily, fontSize, lineHeight, letterSpacing, padding, borderRadius, backgroundShape) is how drafts drift into a look that matches no preset. Tune `color` and `bounds` per slide, and use the same preset across every slide of one slideshow. - `preview_image_urls`: optional composed slide preview URLs to attach to the saved card; when omitted, the card best-effort harvests one allowlisted source image per page from the first variant's sceneData (not composed renders) - `create_share_link`: defaults to `false`. Set `true` to return an existing link or create a new one; a new link without a password is accessible to anyone who has its URL - `caption`: optional override for the single-scene path; defaults to `scene_data.caption` when present Keep every text block inside the platform safe area: on a 1080x1920 slide that is x 46-1033, y 115-1545, minus the platform UI sitting inside it (x 313-762, y 133-216; x 887-1019, y 960-1092; x 886-1020, y 1152-1316; x 887-1019, y 1358-1521). Text crossing those rects is covered by the app's own chrome once the slide is posted, so it reads as a broken slide however good the copy is. A block grows to fit its text, and a centered block grows around its center, so leave room on both sides: wide copy belongs above the action rail rather than beside it. The write boundary moves a block down or up when that is enough to clear the chrome, but it never rewrites or shrinks copy. Size the text to the room it has. Example `upsert_slideshow_draft` payload that keeps a slideshow editable and reviewable as sceneData: ```json { "draft_name": "musicluvr20 artists i always come back to", "create_share_link": true, "scene_data": { "id": "scene-musicluvr20-20260408-01", "name": "musicluvr20 artists i always come back to", "pages": [ { "id": "page-cover", "name": "Cover", "width": 1080, "height": 1920, "background": "#111111", "blocks": [] } ], "pageOrder": ["page-cover"], "activePageId": "page-cover", "defaultPageSize": { "width": 1080, "height": 1920 }, "templateMode": true, "caption": "5 artists i always come back to" } } ```

upsert_slideshow_draft

Schedule the content agent for the active product: a routine that fires on a recurring schedule (every few hours, daily, or on chosen weekdays) and asks for a batch of posts for the product's accounts each time. For product members, every scheduled time sends the Doublespeed team a request to write the batch; the drafts arrive as review links on the product's Content agent page (workspaceUrl in the result). Doublespeed staff choose per routine with generate_mode: "run" starts a content agent run at each scheduled time and is the staff default, "request" sends the team a request; ask a staff member which one they want when they have not said. Staff may also pin the content agent's pipeline settings for this routine (allow_new_ai_images, allow_image_detail_edits, trend_search_enabled, skip_judge_pass, use_post_learnings, qa_enabled); each one left out follows the product's content agent settings at each scheduled time. Product members leave them out. A product holds at most 5 routines, which are edited, paused, and deleted on the Content agent page. Confirm the schedule and its timezone with the person before calling.

create_content_agent_routine

Rank slideshow copy candidates with a learned scorer trained on this platform's real posted-slideshow outcomes (within-account winner/loser pairs, so account effects are cancelled). Send ALL candidates in ONE call: scores are relative within the submitted set only. Returns raw linear scores and score10 (10 times the mean model-implied pairwise win rate against the other submitted candidates). Small raw margins stay near 5. These predictions have not been calibrated against fixed-age or human outcomes; score10 is not confidence or a quality grade. Higher score10 means the copy pattern historically outperformed sibling posts. It judges copy only (slide texts + caption), knows nothing about brand rules or factual accuracy, and reflects patterns resembling past winners, so pair it with editorial judgment for constraints and novelty.

score_slideshow_copy

Find real reference photos on Pinterest. Returns raw image URLs and dimensions. Download and inspect a chosen photo, then upload it with prepare_media_upload and commit_media_upload before placing it in a draft.

search_pinterest_images

Select the active product used by product-scoped MCP tools. With a session, this changes that session; persist_default: true also saves the default for future sessions. For a sessionless connection, selecting a product saves the default even when persist_default is false. Call list_products to see accessible products and inspect persistedDefault in the result.

set_product

Start the content agent for specific accounts of the active product, the same as the Generate button on the product's Content agent page. The agent drafts posts for each named account from the product wiki and delivers them as review links. Returns at once with the run id: a run takes twenty minutes or more per account and spends the product's generation credits. Each call starts a new run. Progress, drafts, and review links appear on the Content agent page (dashboardUrl in the result); this tool does not return progress on repeat calls. Settings left unset follow the agent's own settings on that page. One post per account unless post_count says otherwise. Staff only: mode "request" sends the brief to the team without running the agent, and allow_image_detail_edits and trend_search_enabled pin those settings for one call.

start_content_agent_run

Update fields on a social media account. Only provide the fields you want to change. Bio has a 160 character limit.

update_account

Change individual items in a saved editor project without resending the whole composition: retrim or move a scene, edit a headline, swap a motion clip's HTML, hide or remove an item, add a text item or motion clip, or replace the captions. Item ids come from the items index returned by save_editor_project (or include_composition). Ops apply in order; the project is re-saved and the open editor picks the change up live. Layout and timing edits are also available to the user directly in /editor; use this tool for content changes an agent makes.

patch_editor_project